[ Up to Semantics: the study of meaning]
EG: What is, say, a dog?
One way of seeing our representation of the 'meaning' of dog, is that we have a large number of exemplars of dogs (taken from our experience), and perhaps in the 'middle' of these exemplars a prototype of a dog that can 'morph' as needed.
Semantics has both an 'internal' and an 'external' aspect. Clearly we ground meanings in our own experiences. If the meanings of words are based on our individual representations, in order for communication to take place, there must be common experiences shared between the members of a language community. People who are inexperienced in advanced calculus may have the feeling they are reading gibberish upon trying to read a textbook in that field. Indeed, it will continue to be gibberish until you've done a large number of the exercises in each chapter.
Many language constructions are analyzable, but not predictable. EG: 'Fire house' vs. 'Dog house'. We can see the rules that motivate each use of 'house', but we have no a priori way of knowing which rule would apply in new constructions.
[ Up to Semantics: the study of meaning]
Meaning for people basically amounts to the images (or smells, sounds, ...) that are conjured in the understander's brain. For computers, semantics is more precise, a much less rich. It can be characterized as the output associated with an input.
A jpeg graphics file is written in a particular format which, when interpreted by an appropriate graphics program results in an image on the screen.
A postscript file is written in a different format, which a laser printer can interpret to produce a printed document. HTML is another example.
In each of these cases, the files are just strings of bits. What gives them meaning is the mesh between the patterns of the bits and the interpreter applied to them.
Computer applications which attempt to deal with the semantics of natural language range from the key-word searches of a search engine to machine translation. An active area of research is that of Machine Understanding, which takes as input say, input from the newswires about terrorist incidents in Latin America, and produces as output a template filled out saying who the perpetrator was, what the nature of the event was, how many casualties there were, etc.
[ Up to Semantics: the study of meaning]
We can represent meanings in terms of a set of interconnected propositions. EG:"A bird is flying" := There exists some x such that (is-a x bird) and (flying x). This is referred to a predicate logic.
This has to be elaborated on with modal operators like 'maybe', and 'possible worlds' in order more fully to accomodate the range of meanings expressable in natural language.
Whatever knowledge representation scheme we use, a semantic system has to be build on an ontology which systematically lists our declarative knowledge about the world.
A semantic system also has to have rules of inference which allow the understander to fill in the gaps between what is said and what is implied.
[ Up to Knowledge representation]
A semantic system which tries to capture the range of meanings expressable in a language has to produce a systematic representation of 'the world'; a kind of encyclopedia called an ontology which fits new expressions into categories the way syntactic rules fit words and phrases into syntactic categories.
Just as our syntactic rules eventually had to ground themselves in 'terminals', i.e. actual words, our ontology has to ground itself in primitive concepts.
Some possible ontological categories which can be used to express primitives:
Note that these categories correspond to common syntactic categories (nouns, adjectives,...). We should not mistake this for a direct correspondence however, because languages provide various means of moving between categories. EG, we can 'nominalize' a verb (see -> sight).
Once we've put a formalism for producing more complex categories out of primitives, we have a 'generative semantics' which like its syntactic counterpart can generate an infinite number of meanings.
As we process texts, we can refer to our ontologies for prototypical knowledge about the things spoken about, but we must also be able to express the fact of specific entities which exist, and which can change form without changing identity. Recall the tin woodman in the Wizard of Oz, who starts out as a human, and because of a curse by the Wicked Witch of the East, his ax successively chops off every piece of his body, successively to be replaced by a tin prosthesis (except of course, his heart).
[ Up to Knowledge representation]
Much of language works by reminding us of patterns or schemas which occur frequently in our lives. EG, 'I bought a newspaper for a dollar' invokes a frequently occuring, but actually fairly complex event, in which there is 1) a buyer ('I'), 2) a seller (unspecified), 3) Some good ('a newspaper'), 4) a price ('a dollar'). When such an expression is used, we can make a number of inferences about the relationships that hold between the fillers of these roles (for example that the newspaper now belongs to me, and the seller is a dollar richer). Schemas like this that extend over some period of time (a restaurant or doctor's office are typical examples) are called 'scripts'.
Sometimes it serves our purposes to try to express the roles that define events as a few general types of thematic roles. There is no standarized set of thematic roles, but some typical ones are:
So for example most common meaning of drink could be defined as:
[agent] causes [patient] to go on [path] from [source] to [destination]
where [agent] is a person, [patient] is some liquid, [source] is some container for the liquid, [destination] is the digestive tract of the [agent], and [path] is a route between [source] and [destination].
Note that the verbs cause and go are serving as two of a limited set of verbal primitives out of which words like drink are being defined. We can put constraints on what kinds of entities can be involved with these verbs, for example 'go' must always involve a [path].
An advantage of using thematic roles is that we can make generalizations about the grammar of the language: typically the [agent] takes the subject position, while a noun phrase following the word 'from' is probably serving as the [source].
Note also that we can give [patient] a default value of 'alchohol', which would fill that role in cases where no object of the verb is specified.
Let's take another example: 'Run' which could be defined as [agent] causes [agent] to go on [path] from [source] to [destination]. This is adequate if all we want to do is track the location of our agents. It does not express the difference between run and walk. Typically can add resolution by adding features, say [+/- fast] or [speed = fast], or [manner = runningly].
[ Up to Knowledge representation]
In addition to expressing a catalog of the kinds of things and events that we come across in what is called declarative knowledge, we need to be able automatically to tie knowledge together allow us to maintain a consistent image of the world. Typically this is characterized as a set of procedural knowledge which takes the form of rules: 'if x is a bird, then x lays eggs'.
We also need to be able to express constraints on which words and ideas can be combined meaningfully.
[ Up to Knowledge representation]
Ask someone a question, they're likely to give you a response quickly.
Schank asserts that this is because they're retrieving a piece of knowledge 'off the shelf', with a little bit a adaptation.
Remindings are the key to intelligence. When encountering a new circumstance we
1) are reflexively reminded of some past event, then
2) analogize the present circumstances to this reminding.
Conversations are often an exchange of remindings amongst the participants.
[ Up to Schank:Knowledge is Stories]
Much of our understanding of the world employs scripts. A script is set of expectations everyone has about certain frequently occurring situations. Often these situations are quite complex.
Some scripts are personal, some are more public.
Typical examples of scripts are restaurants, trips to the doctor's office...
Somehow certain individual instances of these scripts (episodes) are memorable and leap to mind under certain conditions.
[ Up to Schank:Knowledge is Stories]
In order to be reminded about a given event, we have to put it into memory in such a way that it will be retrieved when most appropriate.
An important aspect of indexing seems to revolve around expectation failures. Another is that of goals being pursued or thwarted by the actor in the episode. Why does this make sense?
At the kernal of many stories and episodes in our lives there is some 'gist' which somehow summarizes its importance and applicability. Often this gist is best characterized by the explanation we've come up with for the expectation failure or thwarted goal.
[ Up to Schank:Knowledge is Stories]
This whole process requires that we be able to analogize between the episode retrieved, and the situation at hand.
Central to this is the idea of roles. Applying an analogy in general usually means identifying corresponding entities which fill the same roles in both cases.
[ Up to Knowledge representation]
Inheritance is a way of declaratively making generalizations about classes and subclass which allow us to follow links in the heirarchy as a means of making inferences.
We have already gone into simple inheritance for categories of things.
Jacobs and Rao describe a way of defining a very powerful inheritance heirarchy for events and roles. Because it uses inheritance, it allows us to make the same generalities that pertain to thematic roles, but also allows us to express events of greater complexity than can be done straightforwardly with basic case frames involving thematic roles, such as buying and selling.
The root node for an event has a basic category of role [participant].
Simple events such as transferal can be characterized with roles like [object] [source], [destination] which can be said to inherit from [participant]
In addition to simple events, we can describe complex events, which are characterized by [subevent]s.
So we can define these transfer events:
We can define a complex event 'commercial transaction', with roles:
Having defined a schema for commercial transactions, we can define words like 'buy' and 'sell' which refer to the various parts of this construct in their lexical definitions: [customer] buys [goods]. Note that we can infer from the commercial transaction's structure that whatever follows '...buy...from...' should be the [merchant], since [goods] inherits unambiguously from the [delivery] sub-event, in which [merchant] inherits from [source].
We don't have to stop there, we can define a 'go to supermarket' event, and define [owner], [manager], and [cashier] inherited from the [merchant] role, define more specific sub-events which typically apply, etc.
[ Up to Knowledge representation]
When we make these definitions, we are defining classes of word forms and classes of objects in the real world to which we ascibe various generalities. In the course of dealing with actual reality, we encounter instances of these classes. When we encounter something and assign it to a class, we bestow upon that instance all or at least most of the generalities we associate with the class. Sometimes this is referred to as the type/token distinction. Types are abstract, tokens are actual things.